The main goal of manifold learning is to find a smooth low-dimensional manifold embedded in high-dimensional data space. At present, manifold learning has become a hot issue in the field of machine learning and data mining. In order to seek valuable information from high-dimensional data stream and large-scale data set, it is urgently necessary to incrementally find intrinsic low-dimensional manifold structure in such observation data set. But, current manifold learning algorithms have no incremental ability and also can not process the giant data set effectively. Aiming at these problems, the concept of incremental manifold learning is firstly defined systematically in this paper. It is advantageous to interpret the dynamic process of developing a stable perception manifold and to guide the research of manifold learning algorithms which fit to incremental learning mechanism in man brain. According to the guiding principles of incremental manifold learning, a dynamically incremental manifold learning algorithm is then proposed, which can effectively process the increasing data sets and the giant data set sampled from the same manifold. The novel method can find the global low-dimensional manifold by integrating the low-dimensional coordinates of different neighborhood observation data sets. Finally, the experimental results on both synthetic “Swiss-roll” data set and real face data set show that the algorithm is feasible.